chore: adapt path and training hparams

2026-05-09 21:40:10 +08:00
parent 75eacf00c2
commit 46560f5740
6 changed files with 32 additions and 30 deletions
--- a/scripts/kaiyuan2b-training/params/qwen3_1p7b/data_phase1_smoke.sh
+++ b/scripts/kaiyuan2b-training/params/qwen3_1p7b/data_phase1_smoke.sh
@@ -1,8 +1,8 @@

-DATA_DIR=/ssd/yi/converted_data/megatron_phase1
+DATA_DIR=/ssd1/yi/converted_data/phase1

 START=0
-END=210
+END=0

 DATA_PATHS=""
 for idx in $(seq -f "%05g" $START $END); do
@@ -13,7 +13,7 @@ DATA_ARGS="
  --data-path ${DATA_PATHS}
  --split 999,1,0
  --tokenizer-type HuggingFaceTokenizer
-  --tokenizer-model /apps/yi/model_training/data/tokenizer
+  --tokenizer-model /ssd1/yi/data/tokenizer
 "

 #   --vocab-file /apps/yi/model_training/data/tokenizer/vocab.json
--- a/scripts/kaiyuan2b-training/params/qwen3_1p7b/hparams.sh
+++ b/scripts/kaiyuan2b-training/params/qwen3_1p7b/hparams.sh
@@ -1,5 +1,5 @@
 HPARAMS="
-  --micro-batch-size 16
+  --micro-batch-size 4
  --global-batch-size 2048
  --train-iters 87000
  --eval-iters 10